They want a tool that makes this file share talk to this SharePoint site which updates this ERP tool over there. The LLM approach is great for the departmental person (if they can still host shadow IT) but falls down at the organizational level. The nature of this work is fundamentally different, crappier, and less interesting than what any person on HN wants to be doing which is a contributor to misunderstanding of the market.
EDIT: fixed grammar.
I wrote a short post about it on my blog: https://blog.waleson.com/2022/10/access2mendix.html
In my most hopeful of futures, we've figured out how to do lightweight inference, and if the models don't run locally at least they aren't harming the planet, and all this AI tooling hydrates all the automation projects of the last 40 years so that my favorite tiny local music label can have a super custom online shop that works exactly the way they need without having to sacrifice significant income to do it.
A lot of value indeed, but not just for less technical people. Imagine ddd vs gdb. Usually some kind of visual debugging aid isn’t available in an environment because the ROI isn’t there, not because technical people love mental parsing or hate graphics. The LLM revolution is changing the calculus here: creating new tools and new visualizations is easier than ever. It would be unthinkable three years ago to create a visual debugging aid just to use it once, outside of truly gnarly and show-stopping bugs; now it could very well be feasible.
Yes. In the GIS industry for example, nothing has fundamentally changed with the introduction of LLMs. They may make the same processes more efficient e.g. though automated building of workflows. AI has significantly improved classification work of course but it's still using the same principles (we've been doing ML longer than most industries). Geocoding will get cheaper and easier but it's still geocoding.
GIS software allowing standard visualisation, export and map production will bet a lot better because of LLMs. It's an area where the sheer complexity and number of formats was overwhelming, but now a GeoTiff parser can be built in a day or two.
The article was making a bit of a sweeping statement based on a single datapoint: they didn't need Retool anymore.
There's a lot of value in having direct manipulation and visual introspection of UIs, data, and logic. Those things allow less technical people to understand what the agents are creating, and ask for help with more specific areas.
The difficulty in the past has been 1) the amount of work it takes to build good direct manipulation tools - the level of detail you need to get to is overwhelming for most teams attempting it - but LLMs themselves make this a lot easier to build, and 2) what to do when users hit the inevitable gaps in your visual system. Now LLMs fill these gaps pretty spectacularly.